People use the word “plugin” for several different things. In one context it means a bundle of reusable instructions and reference files. In another, it may refer to an app or connector that gives an AI client access to tools or data. The label alone does not tell you what the package can do, so I start by asking what it contains and what permissions it needs.
A skills-based plugin can teach an assistant a repeatable way to handle a kind of request. A small consultancy might include guidance for preparing a discovery call, checking which facts are missing, and drafting a follow-up for a person to review. The instructions can include the firm's approved language and process. They do not, by themselves, let the assistant read a CRM, send email, or inspect a calendar.
A connected app can provide a different capability. It may expose a search or action through a defined interface, sometimes using MCP. That is where access to current records or operations can enter. The server and its permissions matter. A plugin package and an MCP server are related building blocks, but they are not interchangeable: instructions tell the model how to approach the work; tools let it request information or actions from another system.
Would your team use it twice?
That is a useful test. If staff repeatedly explain the same standards, offer, or decision process to an assistant, a reusable instruction package can save setup effort and reduce inconsistency. If the task depends on live data, guidance alone will go stale. A connected tool may help, provided there is a reliable source and a clear permission boundary.
Take an incoming service enquiry. A guidance-only package could help the owner identify missing details and draft thoughtful questions from information pasted into the conversation. A connected version might look up an approved service list. It should not promise a delivery date just because a model found an old note, and it should not submit a quote without an authorised person's review.
For this reason, I would avoid beginning with a feature list. Observe the task first. Where do people repeat themselves? Which facts change? What needs a human judgement? A plugin is useful when it fits those answers and lives in an AI client the team can actually access. Client setup, naming, permissions, and the visible interface differ, so instructions should not imply one universal screen or workflow.
Keep the package honest
We have been shaping a Brownsmith business-improvement toolkit around guided workflows for diagnosis, workflow review, customer acquisition, and implementation planning. Its instructions help the assistant ask for context and separate known facts from assumptions. That is distinct from our MCP work, which concerns bounded access to tools. You can read our explanation of when a service business might need MCP.
If you're considering a tailored assistant workflow, Brownsmith Dynamics' business-efficiency service starts with the job to improve. Send a note on WhatsApp or through the contact page. A short conversation can establish whether you need a reusable guide, a system connection, or simply a clearer process.